Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2021 Mar 19;13(6):1415.
doi: 10.3390/cancers13061415.

Fine-Tuning Approach for Segmentation of Gliomas in Brain Magnetic Resonance Images with a Machine Learning Method to Normalize Image Differences among Facilities

Satoshi Takahashi  1   2 , Masamichi Takahashi  3   4 , Manabu Kinoshita  5 , Mototaka Miyake  6 , Risa Kawaguchi  7 , Naoki Shinojima  8 , Akitake Mukasa  8 , Kuniaki Saito  9 , Motoo Nagane  9 , Ryohei Otani  10   11 , Fumi Higuchi  10 , Shota Tanaka  12 , Nobuhiro Hata  13 , Kaoru Tamura  14 , Kensuke Tateishi  15 , Ryo Nishikawa  16 , Hideyuki Arita  5 , Masahiro Nonaka  17   18 , Takehiro Uda  19 , Junya Fukai  20 , Yoshiko Okita  18   21 , Naohiro Tsuyuguchi  19   22 , Yonehiro Kanemura  18   23 , Kazuma Kobayashi  1   2 , Jun Sese  7   24 , Koichi Ichimura  4 , Yoshitaka Narita  3 , Ryuji Hamamoto  1   2
Affiliations

Fine-Tuning Approach for Segmentation of Gliomas in Brain Magnetic Resonance Images with a Machine Learning Method to Normalize Image Differences among Facilities

Satoshi Takahashi et al. Cancers (Basel). .

Abstract

Machine learning models for automated magnetic resonance image segmentation may be useful in aiding glioma detection. However, the image differences among facilities cause performance degradation and impede detection. This study proposes a method to solve this issue. We used the data from the Multimodal Brain Tumor Image Segmentation Benchmark (BraTS) and the Japanese cohort (JC) datasets. Three models for tumor segmentation are developed. In our methodology, the BraTS and JC models are trained on the BraTS and JC datasets, respectively, whereas the fine-tuning models are developed from the BraTS model and fine-tuned using the JC dataset. Our results show that the Dice coefficient score of the JC model for the test portion of the JC dataset was 0.779 ± 0.137, whereas that of the BraTS model was lower (0.717 ± 0.207). The mean Dice coefficient score of the fine-tuning model was 0.769 ± 0.138. There was a significant difference between the BraTS and JC models (p < 0.0001) and the BraTS and fine-tuning models (p = 0.002); however, no significant difference between the JC and fine-tuning models (p = 0.673). As our fine-tuning method requires fewer than 20 cases, this method is useful even in a facility where the number of glioma cases is small.

Keywords: MR images; deep learning; fine-tuning; glioma; machine learning.

PubMed Disclaimer

Conflict of interest statement

No potential conflicts of interest were disclosed by any author in this study.

Figures

Figure 1
Figure 1
The overall flow of this research. Three types of machine learning models were built for segmentation: The Multimodal Brain Tumor Image Segmentation Benchmark (BraTS) model, the Japanese cohort (JC) model, and the fine-tuning model. (1) The BraTS model was trained in the training part of the BraTS dataset. (2) The BraTS model was evaluated by the Dice coefficient scores of the BraTS data set and the test portion of the JC data set. (3) The JC model was trained on the training portion of the JC data set. (4) The JC model was evaluated by the Dice coefficient score of the test portion of the JC data set. (5) The BraTS model was fine-tuned for optimal analysis at each facility. (6) The BraTS model fine-tuned by the aforementioned method, was named the fine-tuning model. (7) The fine-tuning model was evaluated with the Dice coefficient score of the test portion of the JC data set.
Figure 2
Figure 2
The results of three types of machine learning models for segmentation. A bar graph shows the Dice coefficient score for each facility. The horizontal axis is the facility, the vertical axis is the Dice coefficient score, and the colors indicate the types of machine learning models for segmentation. The Dice coefficient score of the fine-tuning model was significantly improved compared to that of the BraTS model and was comparable to that of the JC model.
Figure 3
Figure 3
The results of three types of machine learning models for segmentation focused on pathological diagnosis. A bar graph shows the Dice coefficient score for each pathological diagnosis. The horizontal axis shows the pathological diagnosis, the vertical axis shows the Dice coefficient score, and the colors show the type of machine learning models for segmentation.
Figure 4
Figure 4
Segmentation results on Case 3. The left column is ground truth (made by a skilled radiologist), the middle is predicted by the BraTS model, and the right is predicted by fine-tuning model specialized for the facility has Case 3. The volume of interest (VOI) predicted by the BraTS model had an area that didn’t seem to be a brain tumor. However, that by fine-tuning model specialized for the facility has Case 3 only have an area that seemed to be a tumor.
Figure 5
Figure 5
The results of comparison of image types in Case 1 from the BraTS dataset and Case 2 and Case 3 cases from the JC dataset are shown in histograms. The horizontal axis of the histogram represents the voxel value converted into a Z-score, and the vertical axis represents the number of voxels. (A) Each color represents image types. The GdT1 part of histograms of Case 1 and Case 2 have one peak around 1.8, however that of Case 3 has no peak. (B) The image histograms show the relationship between the T2 image (equivalent to the slice corresponding to the VOI) and true VOI. Blue histogram represents the T2 image, and orange represents true VOI. The VOI part of histograms of Case 1 and Case 2, located right edge (around 3), occupied a relatively large portion. But that of Case 3 was located central (around 2.5) and had a small portion.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

See all "Cited by" articles

References

    1. Rasmussen B.K., Hansen S., Laursen R.J., Kosteljanetz M., Schultz H., Nørgård B.M., Guldberg R., Gradel K.O. Epidemiology of glioma: Clinical characteristics, symptoms, and predictors of glioma patients grade I–IV in the the Danish Neuro-Oncology Registry. J. Neuro-Oncology. 2017;135:571–579. doi: 10.1007/s11060-017-2607-5. - DOI - PubMed
    1. Arita H., Narita Y., Fukushima S., Tateishi K., Matsushita Y., Yoshida A., Miyakita Y., Ohno M., Collins V.P., Kawahara N., et al. Upregulating mutations in the TERT promoter commonly occur in adult malignant gliomas and are strongly associated with total 1p19q loss. Acta Neuropathol. 2013;126:267–276. doi: 10.1007/s00401-013-1141-6. - DOI - PubMed
    1. Ichimura K., Nishikawa R., Matsutani M. Molecular markers in pediatric neuro-oncology. Neuro-Oncology. 2012;14:iv90–iv99. doi: 10.1093/neuonc/nos204. - DOI - PMC - PubMed
    1. Hegi M.E., Diserens A.-C., Gorlia T., Hamou M.-F., De Tribolet N., Weller M., Kros J.M., Hainfellner J.A., Mason W., Mariani L., et al. MGMTGene Silencing and Benefit from Temozolomide in Glioblastoma. N. Engl. J. Med. 2005;352:997–1003. doi: 10.1056/NEJMoa043331. - DOI - PubMed
    1. Lassman A.B., Iwamoto F.M., Cloughesy T.F., Aldape K.D., Rivera A.L., Eichler A.F., Louis D.N., Paleologos N.A., Fisher B.J., Ashby L.S., et al. International retrospective study of over 1000 adults with anaplastic oligodendroglial tumors. Neuro-Oncology. 2011;13:649–659. doi: 10.1093/neuonc/nor040. - DOI - PMC - PubMed

LinkOut - more resources